I think the twist axis was facing wrong in the beginning because the bone orientation was not aligned. Maybe I can just ignore the joint order entirely? (I did not yet quite understand if the joint order really does matter for anything other than intuitive Naming of the rotation parameters?)